V1 open-source scope
This document is the source of truth for the first public GenerativeDynamics release. It prevents experimental research branches from becoming accidental public APIs.
Product promise
GenerativeDynamics is a unified planning and control stack for robotics. V1 lets users run interchangeable planners against shared environments, constraints, metrics, visualizations, and experiment protocols from versioned configuration files.
Included capabilities
- Config-driven experiment execution with plugin registries for methods, environments, metrics, obstacle generators, and visualizations.
- JAX-first accelerated planning and control on supported CPU and GPU devices.
- Planning methods: MDOC, MD-COAS, 2GO, MGA, MBD, EB-MBD, MPPI, DIAL-MPC, ATACOM, ISSA, PegasusFlow, standalone RL, and model-based-only comparisons.
- Specialized integrations: DPCC and SafeDiffuser when their optional Torch dependencies are installed.
- Reproducible runs with explicit seeds, resolved configurations, metrics, and saved artifacts.
- Robot task recipes for planar navigation, D3IL avoiding, quadruped stepping stones, humanoid corridor navigation, peg insertion, and surface scanning.
MDOC, MD-COAS, 2GO and MGA are our algorithms. The planner catalogue lists each algorithm with a short description and its source.
MD-COAS is the published algorithm name. MDCOAS is its historical compact
spelling, and cfsmbd / cfsmbd_full are historical implementation keys for
the same algorithm. CFS-MBD is not a separate algorithm in the public
catalogue. Existing filenames containing mdcoas-a or mdcoas-f retain
configuration settings for reproducibility; they are not additional named
planners. 2go and twogo identify the same method; dial is the DIAL-MPC
implementation key.
Deferred from V1
- Soft-robot simulation and morphology co-design.
- 3D Gaussian Splatting and active scene reconstruction.
- MGA GPU readiness claims.
- Humanoid walking and push-to-line task readiness claims.
- General Torch planner parity and Rust/C++ execution backends. V1 already contains a Torch runtime helper and dedicated Torch integrations; that does not establish Torch support for the JAX planner catalogue.
Soft-robot, co-design, and 3DGS work must not be imported, registered, documented as available, or installed through the default V1 package. MGA GPU and humanoid push remain visible only as deferred development work until their release gates pass.
Release task matrix
| Task | Configuration root | V1 status |
|---|---|---|
| Planar navigation | configs/single_2d |
Included |
| D3IL avoiding | configs/d3il_avoiding |
Included |
| Quadruped stepping stones | configs/quadruped/stepping_stones_2d |
Included |
| Humanoid corridor | configs/humanoid/corridor_2d |
Included |
| Manipulator peg insertion | configs/arm/peg_insert |
Included |
| Manipulator surface scan | configs/arm/surface_scan |
Included |
| Humanoid push-to-line | configs/humanoid/push_to_line |
Deferred |
Public release gates
V1 is ready to tag only when all of the following are true:
- A clean environment can install the core package and at least one supported simulator extra.
- Every included recipe has one documented command and a deterministic smoke test.
- CI covers package build, installation, import, unit tests, configuration validation, and the CPU smoke path.
- The README states measured capabilities and limitations without promising deferred work.
- License, citation, contribution, security, and code-of-conduct files are complete.
- Published benchmark numbers include hardware, software versions, warm-up, compilation treatment, batch size, horizon, seeds, and raw result artifacts.